Release Notes v0.1.0

Date: 2026-08-20
Package: planner-critic v0.1.0 — PyPI

What's New

Hierarchical task planning — decompose a goal into a typed PlanVersion with tasks, dependencies, branches, preconditions, and verification/rollback steps
Independent LLM critic — audit every plan revision against six heuristic families (feasibility, risk, missing_steps, unsafe_sequencing, unverified_dependencies, weak_rollback) with severity-graded findings
Plan revision loop — planner revises based on critic findings; loop terminates on approval, escalation, budget exhaustion, convergence stall, or revision cap
Risk tolerance — balanced (findings are advisory warnings, gates are the hard floor) and strict (zero tolerance for any finding, fail-closed)
7 deterministic gates — injection-immune gates for ordering-sanity, branch-sanity, rollback-safety, verification-safety, preconditions-referenced, branch-tasks, and high-risk completeness
Escalation management — human-in-the-loop with override, patch, and restart decisions
Convergence detection — early termination when the planner stops making meaningful progress, saving LLM calls
Plan revision policies — patch, restart, abort

CLI — plancritic plan, critique, field-test run, demo, quickstart, migrate, serve
Python API — Engine.plan(), ApprovedPlan, LoopResult, Finding, Escalation
HTTP server — health check, plan, and escalation endpoints
Provider registry — pluggable LLM providers (OpenRouter, OpenAI, oMLX, Ollama) via TOML config
StructuredEnforcer — retry mechanism for LLM JSON output, fail-closed after 3 retries

Field test harness — 157 goals across 35 domains, automated scorecard generation
Field test results — Scorecard A (strict plan semantics) and Scorecard B (pass* semantics) published
Release gate — §7.3 criteria adjudicated: PASS with planned amendments

Metric	Result
Balanced goals approved	71/71 (100%)
Strict goals escalated	81/81 (100%)
Adversarial goals escalated	8/8 (100%)
True failures	0
Deterministic gate passes	156/157 (99%)
Scorecard A (post-amendment)	PASS
Scorecard B (pass* semantics)	100%

None — v0.1.0 is the initial release.

Planner capability gap — 132 concrete blockers across 63 strict goals, concentrated in 3 families (unverified_dependencies, unsafe_sequencing, weak_rollback). A deterministic precondition closer in v0.2.0 would eliminate 48%.
CLI surfaces — demo, quickstart, migrate return non-zero exit codes
HTTP + MCP surfaces — not yet tested
Adapter coverage — only raw Python adapter ran
Multi-model — only gpt-4o-mini tested
Finding quality audit — noise-finding rate not yet measured
Executor usability — not yet audited

Doc	Location
Field Test Results	`docs/field-test/field-test-results-0.1.0.md`
Field Test Plan	`docs/field-test/field-test-plan.md`
Architecture	`docs/architecture/architecture-v0.1.0.md`
API Reference	`docs/reference/api.md`
Design Decisions	`docs/design/design-decisions.md`
Demo Scenario	`docs/design/demo-scenario.md`
Changelog	`CHANGELOG.md`